Purpose: To demonstrate the benefits of kernel density estimators to enhance understanding of why cost-effectiveness can vary by patient subgroup.

Methods: We obtained data from a previously published study showing differential cost-effectiveness for two patient subgroups. Using kernel density estimation, we explored the reasons for the differences in the incremental net benefit and illustrated our findings with a graph of the resulting kernel density estimate (similar in spirit to a probability density function). The net benefit (nb) was calculated as patient-level net monetary benefit (i.e., nb = le – c where l is society's willingness to pay, e is the patient's health outcome and c is the patient's cost). As a sensitivity analysis, various l values were used. The kernel density estimate of nb was formed by summing weighted values of the nb. For the kernel function providing the weights, we chose the Epanechnikov kernel, i.e., 0.75(1 – 0.20z²)5^-0.5 where z = (nb – NB)/h and h = 0.9min(SD_nb, interquartile range/1.349)n^-0.2.

Results: Kernel density estimation is standard in common statistical packages, so producing the kernel density estimators was quite feasible. Overlaying the graphs of the kernel density estimates illustrated that the distributions of nb were different for the “usual care” and “new treatment” patient subgroups both in terms of central tendency and shape. This suggests that the net benefit of “usual care” differs between the patient subgroups. In addition the kernel density estimation results hint that the net benefit of “new treatment” may differ between the two groups. These findings could be easily missed with a conventional analysis of incremental net benefits. In addition to raising the issue of why the differences are occurring, our results also raise the possibility that equity issues may be identified and addressed.

Conclusions: Subgroup analysis is becoming a more relevant aspect of economic evaluation. As its popularity grows, however, it is important for the results to illustrate how differences in incremental cost-effectiveness are achieved. Kernel density estimation can be used to provide insight in this capacity.